TOP 2% Engineering: /PLAN 2026

Detailed Summary

Introduction: The Top 2% Engineer (0:00 - 0:47)

The video opens with a personal goal to become a top 2% agentic engineer by 2026 and outlines a roadmap to achieve this, emphasizing that the insights are based on extensive experience and conversations in the field.

The Era of Powerful Models (0:48 - 3:00)

• Gemini 3 Flash's Impact: Gemini 3 Flash is highlighted as an "insane" model due to its top-tier intelligence, competitive pricing, and lightning-fast speed (200 tokens per second).
• Multimodal Capabilities: It accepts diverse inputs including text, images, audio, video, and PDF.
• Usage Trends: Open Router data shows a 385% token usage increase for Gemini 3 Flash, indicating its rapid adoption and potential to reach top ranks.
• Competitive Landscape: The emergence of Gemini 3 Flash, Opus 4.5, and GPT-5.2 signifies a highly competitive model ecosystem, driving innovation and lowering prices.
• Intelligence per Token: The key takeaway is that intelligence per token is skyrocketing, meaning engineers get more capabilities for less cost.

The Year of Trust (3:01 - 6:00)

• Models are No Longer the Limitation: The speaker asserts that models are no longer the bottleneck; the limitation lies with the engineers themselves.
• Central Thesis: The Year of Trust: 2026 is framed as the "year of trust" in agentic systems, where the central question for top engineers is "how much do you trust your agents?"
• Trust Equals Impact: Trust in agents is directly linked to speed, iteration, and ultimately, impact in engineering work.

Track Record and Opportunity (6:00 - 7:28)

• Prediction Accuracy: The speaker reviews past prediction accuracy: 16/24 in 2023, 36/41 in 2024, and 13/15 in 2025, demonstrating an improving track record.
• Engineer, Not YouTuber: Emphasizes that the speaker is an engineer actively using and betting on this technology.
• Information Advantage: Highlights the importance of preparing for opportunities before they arise to gain a significant information advantage.

Bet 1: Anthropic Becomes a Monster (7:29 - 10:58)

• Masterclass in Execution: Anthropic's consistent execution, beyond initial "luck" with Claude Code and Sonnet 3.5, is praised.
• Engineer-Focused: Anthropic's focus on engineers (tool calling, agent harness, consistent execution) makes it a strong bet for top engineers.
• Foundation for Trust: Betting on Anthropic provides agents with the best foundation, reducing time spent fighting tooling and increasing shipping efficiency.
• Evolution of Anthropic's Offerings: A timeline showcases key releases like Sonnet 3.5, Claude Code, Claude Code SDK (custom agents), sub-agents, and Opus 4.5, demonstrating a shift beyond just models to comprehensive tooling.

Bet 2: Tool Calling is the Opportunity (10:59 - 15:34)

• Low Tool Call Usage: A surprising statistic from Open Router reveals that only 15% of output tokens are tool calls, indicating a massive untapped opportunity.
• Tool Calls as Impact: Tool calls are presented as a proxy for impact, representing agents taking real actions (e.g., calling bash commands, reviewing work, running skills).
• Maximizing Impact: Top engineers will push agents to call longer chains of tools to increase overall impact.
• Trust in Long Sequences: The length of correct tool call sequences directly reflects the level of trust an engineer has in their agents.
• Anthropic's Leadership in Tool Calling: Charts from Open Router show Anthropic models (Claude 3.7, Claude 4 Sonnet, Claude 4.5 Opus) consistently dominating the market for tool-calling capabilities.

Bet 3: Custom Agents Above All (15:35 - 18:43)

• Automating Massive Problems: A custom agent with minimal code (50 lines, 3 tools, 150-line prompt) can automate thousands of hours of manual work.
• Solving Custom Problems: Custom agents are crucial for solving specific, high-value problems that generate significant revenue.
• Table Stakes: Using agentic coding tools is now standard; the next step is building custom agents with SDKs.
• Skyrocketing Trust: Fine-tuning an agent for a specific problem dramatically increases trust in that agent.
• Engineer's Role: The limitation is not the model but the engineer's ability to build the right agent with the correct context, model, prompt, and tools.

Bet 4: Multi-Agent Orchestration (18:44 - 21:30)

• One Agent is Not Enough: Just as one prompt was insufficient in 2023, one agent is now insufficient.
• Scaling Impact: An engineer running 10 agents will significantly outperform one running a single agent, leading to magnitudes of difference in output over time.
• Verification and Trust: More agents provide more verification, directly increasing trust through techniques like cross-validation.
• Sub-Agents as a Start: Claude Code's sub-agents feature was an early indicator of multi-agent orchestration, where a primary agent prompts sub-agents.

Bet 5: Agent Sandboxes (21:31 - 24:40)

• Safe Execution Environment: As agents scale, the need for isolated execution environments (sandboxes) becomes critical.
• Deferring Trust: Sandboxes allow engineers to defer trust; agents can run rampant, and only the successful outcome is considered, similar to dev/staging environments.
• Best of N Pattern: This enables running multiple agents on a problem and selecting the best result, applicable to both greenfield and brownfield codebases.
• Scaling Impact: Sandboxes facilitate parallelization, scaling, and isolation, ultimately increasing impact and deferring trust until the final merge.

Bet 6: In-Loop vs. Out-Loop (24:41 - 27:04)

• Out-Loop Agentic Coding: Involves prompting an external system (e.g., Slack, GitHub) and letting agents autonomously handle the work, delivering a concrete result like a PR.
• In-Loop Agentic Coding: The traditional method of terminal-based, interactive agent use, where the engineer babysits the agent.
• Progressive Offloading: Top engineers will progressively offload more tasks to out-loop systems as trust grows.
• Maximizing Out-Loop, Minimizing In-Loop: The goal is to maximize autonomous tasks and minimize manual intervention.
• Owning Your Systems: Emphasizes building custom agentic systems around one's codebase rather than relying solely on external cloud tools for critical operations.

Bet 7: Agentic Coding 2.0 (27:05 - 30:36)

• Evolution of AI Coding: From AI coding (writing code) to agentic coding (agents engineering and calling tools), the next step is Agentic Coding 2.0.
• Agents That Conduct Agents: This paradigm involves a lead agent delegating, monitoring, and coordinating specialized command-level agents or sub-agents.
• True Orchestration: This is not just parallelization but true orchestration, where the lead agent effectively manages an engineering team of AI agents.
• Trust in the Conductor: Engineers will trust the lead agent (conductor) to deliver the job, building on the principles of multi-agent orchestration and custom agents.

Bet 8: Benchmark Breakdown (30:37 - 32:15)

• Saturated Public Benchmarks: Public benchmarks will become saturated, with all models achieving high scores (90-100%), losing their signal value.
• Private Evaluation Systems: Top engineers will develop private evaluation systems tailored to their specific use cases, providing a competitive edge.
• Measuring What Matters: Companies and engineers will measure models based on their performance in legitimate engineering work and user problem-solving, not generic benchmarks.

Bet 9: UIs vs. Agents (32:16 - 34:10)

• Agents Eating SaaS: Agents are predicted to disrupt Software as a Service (SaaS) companies, especially those with simple UI-on-database applications.
• Agents as the Interface: Many UIs will evolve into prompt interfaces, resembling chat-based interactions.
• Behind-the-Scenes Agents: The best UIs will run multiple agents in the background to accomplish complex tasks more effectively than deterministic code.
• Control and Speed: Top engineers will favor agents for direct, high-speed interaction with services, bypassing slow UI-driven applications.

Bet 10: AGI Hype Dies (34:11 - 35:44)

• Focus on Agents: The AGI (Artificial General Intelligence) and ASI (Artificial Superintelligence) hype will diminish, and the focus will shift to practical applications of agents.
• Marketing Scheme: AGI/ASI is labeled as a marketing scheme that distracts from tangible engineering value.
• Delivering Value: Top engineers will prioritize delivering maximum value with agents rather than chasing abstract AGI promises.

Bonus Bet 11: End-to-End Agentic Engineers Emerge (35:45 - 38:16)

• First Autonomous Systems: A bold prediction that the first end-to-end agentic engineers will emerge, building complete systems that go from prompt to production with zero human review.
• Building the System That Builds the System: The focus shifts from building applications to building agent systems that construct applications.
• Highest Level of Trust: This represents the ultimate expression of trust in agentic systems, achieving full autonomy.

Conclusion: The Year of Trust (38:17 - 40:56)

• Recap of Bets: A summary of how all 10 bets revolve around increasing trust in agentic systems, from choosing the right labs to building complex orchestration.
• Agentic Coding 2.0 Vision: The ultimate vision involves prompting a lead agent that commands specialized sub-agents to plan, build, review, test, and document work.
• Shift in Mindset: Engineers should adopt a mindset where models don't matter as much as effective agent utilization, moving beyond AGI hype to focus on shipping real work.

Review of 2025 Predictions (40:57 - 48:10)

• AI Coding Becomes Standard (Correct): 84% of developers now use AI coding tools.
• Agentic Coding Begins (Correct): Tools started reliably calling tools, driven by Claude Code.
• Evergreen AI Coding Experience (Correct): Claude Code redefined agentic coding, surprisingly in the terminal.
• Cost of Code Declines (Correct): Cost continues to barrel towards zero, exemplified by Gemini 3 Flash.
• Skill Gap Earthquake (Correct): Massive decline in entry-level roles, rise of senior AI orchestrators.
• LLM Model Review: No Wall (Correct): Models pushed past previous limitations, destroying old benchmarks.
• Hyperspecialized LLMs (Correct): Models like Claude Opus and Function Gemma specialize for specific tasks.
• Year of Infinite Memory (Wrong): Context windows improved, but memory remains a challenge.
• Small Language Models on Devices (Correct): Models like 03, Haiku, and Quinn run on various devices.
• OpenAI Remains Number One (Wrong): OpenAI is no longer in the top spot, with Opus 5 and Gemini 3 Flash leading.
• Industry-Breaking Architecture (Correct): Emergence of world models.
• Exponential Slop (Correct): Prediction that describes itself.
• Big Tech Shrinks, Small/Medium Businesses Grow (Mixed/Correct): Big tech cut jobs, while small businesses thrived with smaller teams.
• Data > UX > Benchmarks (Correct): Proprietary data and better UX are more important than raw model intelligence.

Notable Quotes

"Models are no longer the limitation - YOU and I are." "This is the year of trust. Do you trust your agents?" "Trust equals speed. Speed equals iteration and iteration equals impact." "Anthropic becomes a monster." "Tool calling is the opportunity." "Only 15% of output tokens are tool calls according to Open Router." "Custom agents with custom tools are the best way to maximize tool calls and impact." "One well-crafted agent can replace thousands of hours of manual work." "One agent is not enough. Stop running a single cloud code instance. Start running 3, 5, 10, hundreds." "Let your agent run rampid in its own computer." "Agents that conduct agents." "Public benchmarks get saturated. Every model hits 90 to 100%. They no longer truly give you or I signal." "Agents are eating SAS." "Top 2% engineers stop caring about AGI promises and just lock into delivering maximum value with agents." "We see the first end-to-end agentic engineers emerge." "This is the highest possible level of trust we can have in our agents. This is the north star for trusting your agents." "You don't want to focus on building the application anymore. You want to focus on building the system that builds the system." "The cost of code is barreling towards zero." "There is no wall." "The most important thing now is your user data and the relationship you have with your user and access to that data. Then it's your user experience. And then lastly, it's the language model benchmarks and the language model themselves that use to feel your your new agentic experiences."